ncsn-mnist/runner.py at main · dhunstack/ncsn-mnist · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
import numpy as np
import tqdm

import torch.nn.functional as F
import logging
import torch
import os
import shutil
import tensorboardX
import torch.optim as optim
from torchvision.datasets import MNIST
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
from model import CondRefineNetDilated
from torchvision.utils import save_image, make_grid
from PIL import Image

def anneal_dsm_score_estimation(scorenet, samples, labels, sigmas, anneal_power=2.):
    used_sigmas = sigmas[labels].view(samples.shape[0], *([1] * len(samples.shape[1:])))
    perturbed_samples = samples + torch.randn_like(samples) * used_sigmas
    target = - 1 / (used_sigmas ** 2) * (perturbed_samples - samples)
    scores = scorenet(perturbed_samples, labels)
    target = target.view(target.shape[0], -1)
    scores = scores.view(scores.shape[0], -1)
    loss = 1 / 2. * ((scores - target) ** 2).sum(dim=-1) * used_sigmas.squeeze() ** anneal_power

    return loss.mean(dim=0)

class Runner():
    def __init__(self, args, config):
        self.args = args
        self.config = config

    def get_optimizer(self, parameters):
        if self.config.optim.optimizer == 'Adam':
            return optim.Adam(parameters, lr=self.config.optim.lr, weight_decay=self.config.optim.weight_decay,
                              betas=(self.config.optim.beta1, 0.999), amsgrad=self.config.optim.amsgrad)
        else:
            raise NotImplementedError('Optimizer {} not understood.'.format(self.config.optim.optimizer))


    def train(self):
        tran_transform = test_transform = transforms.Compose([
            transforms.Resize(self.config.data.image_size),
            transforms.ToTensor()
        ])

        dataset = MNIST(os.path.join(self.args.run, 'datasets', 'mnist'), train=True, download=True,
                        transform=tran_transform)
        test_dataset = MNIST(os.path.join(self.args.run, 'datasets', 'mnist_test'), train=False, download=True,
                                transform=test_transform)

        dataloader = DataLoader(dataset, batch_size=self.config.training.batch_size, shuffle=True, num_workers=4)
        test_loader = DataLoader(test_dataset, batch_size=self.config.training.batch_size, shuffle=True,
                                 num_workers=4, drop_last=True)

        test_iter = iter(test_loader)
        self.config.input_dim = self.config.data.image_size ** 2 * self.config.data.channels

        tb_path = os.path.join(self.args.run, 'tensorboard', self.args.doc)
        if os.path.exists(tb_path):
            shutil.rmtree(tb_path)

        tb_logger = tensorboardX.SummaryWriter(log_dir=tb_path)
        score = CondRefineNetDilated(self.config).to(self.config.device)

        score = torch.nn.DataParallel(score, device_ids=[1])

        optimizer = self.get_optimizer(score.parameters())

        if self.args.resume_training:
            states = torch.load(os.path.join(self.args.log, 'checkpoint.pth'))
            score.load_state_dict(states[0])
            optimizer.load_state_dict(states[1])

        step = 0

        sigmas = torch.tensor(
            np.exp(np.linspace(np.log(self.config.model.sigma_begin), np.log(self.config.model.sigma_end),
                               self.config.model.num_classes))).float().to(self.config.device)


        for epoch in range(self.config.training.n_epochs):
            for i, (X, y) in enumerate(dataloader):
                step += 1
                score.train()
                X = X.to(self.config.device)
                X = X / 256. * 255. + torch.rand_like(X) / 256.

                labels = torch.randint(0, len(sigmas), (X.shape[0],), device=X.device)
                loss = anneal_dsm_score_estimation(score, X, labels, sigmas, self.config.training.anneal_power)

                optimizer.zero_grad()
                loss.backward()
                optimizer.step()

                tb_logger.add_scalar('loss', loss, global_step=step)
                logging.info("step: {}, loss: {}".format(step, loss.item()))

                if step >= self.config.training.n_iters:
                    return 0

                if step % 100 == 0:
                    score.eval()
                    try:
                        test_X, test_y = next(test_iter)
                    except StopIteration:
                        test_iter = iter(test_loader)
                        test_X, test_y = next(test_iter)

                    test_X = test_X.to(self.config.device)
                    test_X = test_X / 256. * 255. + torch.rand_like(test_X) / 256.

                    test_labels = torch.randint(0, len(sigmas), (test_X.shape[0],), device=test_X.device)

                    with torch.no_grad():
                        test_dsm_loss = anneal_dsm_score_estimation(score, test_X, test_labels, sigmas,
                                                                    self.config.training.anneal_power)

                    tb_logger.add_scalar('test_dsm_loss', test_dsm_loss, global_step=step)

                if step % self.config.training.snapshot_freq == 0:
                    states = [
                        score.state_dict(),
                        optimizer.state_dict(),
                    ]
                    torch.save(states, os.path.join(self.args.log, 'checkpoint_{}.pth'.format(step)))
                    torch.save(states, os.path.join(self.args.log, 'checkpoint.pth'))

    def anneal_Langevin_dynamics(self, x_mod, scorenet, sigmas, n_steps_each=100, step_lr=0.00002):
        images = []

        with torch.no_grad():
            for c, sigma in tqdm.tqdm(enumerate(sigmas), total=len(sigmas), desc='annealed Langevin dynamics sampling'):
                labels = torch.ones(x_mod.shape[0], device=x_mod.device) * c
                labels = labels.long()
                step_size = step_lr * (sigma / sigmas[-1]) ** 2
                for s in range(n_steps_each):
                    images.append(torch.clamp(x_mod, 0.0, 1.0).to('cpu'))
                    noise = torch.randn_like(x_mod) * np.sqrt(step_size * 2)
                    grad = scorenet(x_mod, labels)
                    x_mod = x_mod + step_size * grad + noise
                    # print("class: {}, step_size: {}, mean {}, max {}".format(c, step_size, grad.abs().mean(),
                    #                                                          grad.abs().max()))

            return images


    def test(self):
        states = torch.load(os.path.join(self.args.log, 'checkpoint.pth'), map_location=self.config.device)
        score = CondRefineNetDilated(self.config).to(self.config.device)
        score = torch.nn.DataParallel(score, device_ids=[1])

        score.load_state_dict(states[0])

        if not os.path.exists(self.args.image_folder):
            os.makedirs(self.args.image_folder)

        sigmas = np.exp(np.linspace(np.log(self.config.model.sigma_begin), np.log(self.config.model.sigma_end),
                                    self.config.model.num_classes))

        score.eval()
        grid_size = 5

        imgs = []

        samples = torch.rand(grid_size ** 2, 1, 28, 28, device=self.config.device)
        all_samples = self.anneal_Langevin_dynamics(samples, score, sigmas, 100, 0.00002)

        for i, sample in enumerate(tqdm.tqdm(all_samples, total=len(all_samples), desc='saving images')):
            sample = sample.view(grid_size ** 2, self.config.data.channels, self.config.data.image_size,
                                    self.config.data.image_size)

            image_grid = make_grid(sample, nrow=grid_size)
            if i % 10 == 0:
                im = Image.fromarray(image_grid.mul_(255).add_(0.5).clamp_(0, 255).permute(1, 2, 0).to('cpu', torch.uint8).numpy())
                imgs.append(im)

            save_image(image_grid, os.path.join(self.args.image_folder, 'image_{}.png'.format(i)))
            torch.save(sample, os.path.join(self.args.image_folder, 'image_raw_{}.pth'.format(i)))

        imgs[0].save(os.path.join(self.args.image_folder, "movie.gif"), save_all=True, append_images=imgs[1:], duration=1, loop=0)